Stem Separation tool for clean up of audio samples


I recently found this android app called which is designed to reverse engineer a final mix into separate stems.

You can find it easily enough online.

It’s kind of interesting for doing remixes I suppose, because it can isolate vocals, drums, bass and lead instruments into individual tracks that you can then export and work with in your own project. If you like that sort of thing.

Personally I haven’t quite decided whether I’m fully down with the ethics of this stuff, but if nothing else I suppose it’s a useful way to solo an element of a tune in order to recreate it with your own sounds… I guess the moral debate about this sort of technology is raging along and obviously pretty thorny.

Putting all that aside for a minute, I’ve been messing about for a bit and by far the most interesting use I’ve found for it is to process samples from movies or TV shows.

For example.

If you upload a screen capture from a horror film for instance, you can process it using a 4 track music algorithm (Vocals, Bass, Drums, Other) and you get some pretty cool results. Generally you’ll end up with tracks split out something like this…

1 “Vocal” track of very clean dialogue.
1 “Drum” track which will mostly be sound effects like gunshots, breaking glass, punches, etc.
1 “Bass” track of low end stuff like explosions, rumbles from collisions, etc.
1 “Other” track, for all the little orchestral stabs or runs or similar melodic atmos stuff.

I mostly enjoy making Jungle-ish tunes so finding interesting vox and FX elements is pretty important to tie everything together, and a big part of the fun for me!

This approach really helps to cut out a lot of the filtering and noise reduction work it usually takes to prep samples of this nature. Particularly useful if you’re only using the tracker to make your music.

The free version lets you process 5 uploads a month and download the results.

I should probably clarify that I have absolutely no affiliation with this company and haven’t done any research into them. Apologies if it turns out they’re working for the devil or something! I’m sure there must be other alternatives out there, maybe even open source, so if anyone knows more then chime in.

Anyway, hope this is interesting and useful.


1 Like

Nice! I often use the stem splitter tool builtin to Koala Sampler to do something similar.


I’ve got Koala Sampler but I’ve not used it much. Certainly never found the stem splitter tool! I’m downloading the neural network thingy now so I’ll compare the results.

Thanks for the tip

1 Like

I’m guessing the AKAI MPC one might become interesting at some point as well (once they release it), since it’s going to be part of their software offering.

Ok, so I’ve just run the same sample through the stem splitter in Koala Sampler and it’s different but still very good.

There definitely seems to be a bit more bleed between the stems in KS and more obvious artifacts. Kind of swooshy, ring mod style stuff going on here and there too… To be fair though, I quite like that sound!

The KS results are a lot louder than the M.a. ones though so it’s possible they’re being normalised or maybe the M.a. results are knocked back a bit.

I’ll have a proper play again once I’m at home and maybe post some results for comparison.

Yeah that MPC One looks cool. Nice touch adding stem separation to it